Research map / Architectures and efficiency / RNNs and normalization
Scaling Laws for SAE Training Data
Nikita Koryagin, Nikita Balagansky, Daniil Gavrilov
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that sparse autoencoder training quality depends mainly on the number of unique activations rather than fresh tokens, so small replayed activation buffers can match much larger training budgets.
Problem
SAE training is often bottlenecked by activation storage, and it is unclear how much stored activation data is actually needed relative to the total training budget.
Approach
Quality depends on the count of unique activations rather than whether every training token is fresh; after a short diversity-limited regime, replaying a fixed buffer preserves most quality, so the buffer can be far smaller than the training budget.
Key results
- Quality of SAE training depends mainly on the number of unique activations rather than on whether every training token is fresh (per abstract)
- After a short diversity-limited regime, additional fresh activations provide little benefit; replaying the same buffer preserves most quality (per abstract)
- Stored activation buffer can be far smaller than the total SAE training budget (per abstract)
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase unique activations from 1M to 128M under a fixed 128M budget (ablation) | BatchTopK SAE 16k · 128M budget | explained variance: 0.5801 -> 0.6393 | Improved |
| increase unique activations from 1M to 128M (ablation) | BatchTopK SAE 32k · 128M budget | explained variance: 0.5870 -> 0.6610 | Improved |
| increase unique activations beyond the plateau (replay instead of fresh data) (ablation) | BatchTopK SAE · fixed budget | explained variance: within 0.005 of max | No clear effect |
| double dictionary size (ablation) | BatchTopK SAE · 1.28B tokens | saturation interval: 20M-40M -> 40M-80M | Hurt |
| replace BatchTopK with TopK or JumpReLU (ablation) | SAE · D=32k, 640M | saturation point: 40M for all | No clear effect |
| math text instead of FineWeb (ablation) | SAE · FineMath-3plus | saturation buffer: 80M, 8x reduction | Mixed |
| Python code instead of FineWeb (ablation) | SAE · Stack-Edu Python | saturation buffer: 40M, 16x reduction | Mixed |
| tighten EV tolerance from 0.005 to 0.0025 (ablation) | SAE · Stack-Edu Python | storage reduction: 16x -> 8x | Hurt |
| let repetition scale depend on dictionary size (ablation) | scaling law · 96-run grid | AIC: delta AIC -77.3 | Improved |
| increase buffer beyond EV-chosen size (ablation) | SAE · SAEBench | AutoInterp, sparse probing | No clear effect |
| increase unique activations (ablation) | SAE 32k and 64k · SAEBench | absorption fraction | Hurt |
Surprising findings
- Replaying the same activation buffer preserves most quality, so fresh activations are largely unnecessary after a short diversity-limited phase
Closest papers by meaning
- Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.90)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.88)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.87)
- Sparse Autoencoders, Again? (ICML 2025 · similarity 0.87)
- Efficient Dictionary Learning with Switch Sparse Autoencoders (ICLR 2025 · similarity 0.86)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
